CN110868359B - Network congestion control method - Google Patents

Abstract

The invention provides a network congestion control method, setting a token bucket in a buffer area before a sending end sends streaming data, sending a credit request to a receiving end, receiving a credit packet sent by the receiving end, increasing the capacity of the token bucket in the buffer area according to the credit packet, and sending data when the size of a data packet to be sent is smaller than the capacity of the token bucket; the receiving end receives a credit request sent by the sending end and sends a credit packet to the sending end; receiving a data stream sent by a sending end; the switch end selectively discards the credit packets according to the link capacity limit, and ECN marks the credit packets which are not discarded according to the proportion of the discarded credit packets. Compared with the prior art, the technical scheme provided by the invention has the advantages that the credit data packet is scheduled with 1538 tokens instead of one data packet, so that the waste of credit during the scheduling of the data packet can be avoided, and the credit loss rate can be effectively calculated by the credit rate control mechanism based on the ECN (engineering-core network) compared with the feedback control based on the credit of the ExpressPass.

Description

A Network Congestion Control Method

technical field

The invention relates to the field of network communication, in particular to a network congestion control method based on data center credit.

Background technique

The infrastructure of data center networks (DCNs) is rapidly evolving, including link speeds, network interface card (NIC) performance, and switch capabilities. However, network users and traffic are also growing, which inevitably leads to network congestion. Severe congestion may cause a sharp drop in network performance, thus requiring an effective congestion control scheme.

At present, many congestion control strategies for controlling network congestion have been proposed, and they can be divided into reactive congestion control and active congestion control. Reactive congestion control aims to control what happens after congestion has occurred. Reactive congestion control typically requires at least one round-trip time (RTT) to cope with congestion and many RTTs to converge to fairness. As link capacity increases rapidly, more and more streaming will be done in a small number of RTTs. Therefore, the reactive congestion control scheme is difficult to meet the requirements of high-speed DCN.

To keep pace with accelerating networks, proactive congestion control schemes are becoming more and more popular. Proactive congestion control schemes such as ExpressPass can avoid congestion and achieve near-zero queuing, very fast convergence, etc. Therefore, active congestion control mechanisms have attracted extensive attention in recent years. However, current proactive congestion control schemes seldom consider workloads with a large number of very short flows, which increases the flow completion time (FCT) of short flows because transport protocols provide latency for short flows that is much higher than the hardware potential, Especially under high network load, for various congestion control schemes, they are based on heavy-tailed workloads where 100KB messages are considered short, such as the Web Search workload used to evaluate DCTCP and pFabric. Unfortunately, in current commercial data centers, there are some applications with much smaller message sizes than Web Server. As a result, existing data center transport designs rarely achieve low latency for minute traffic under high network load.

Also, this issue exists in ExpressPass as well. ExpressPass uses a 5% rate limit on credits based on the minimum and maximum Ethernet frame size, based on the size ratio of credit packets (84 bytes) and data packets (1538 bytes), respectively. However, since short flows (less than or equal to a few hundred bytes) are much smaller than normal packets, the rate-limiting mechanism reduces efficiency and then improves FCT significantly, especially in workloads where the majority of flows are small Down. Moreover, for tail packets with long traffic, there is also such a problem. Since tail latency (99th percentile) is the most important metric for data center applications, existing designs are suboptimal for tail latency under high network load. Therefore, whether the traffic is very short or heavy, an appropriate proactive approach is required for various workloads.

Contents of the invention

To address the above issues, this paper proposes a network congestion control method, which is an active transmission control protocol designed for short flows to achieve low latency in data center networks.

First of all, the present invention provides a network congestion control method, which is applied to the sending end, specifically:

Set the token bucket in the buffer before sending the flow data; send a credit request to the receiving end; receive the credit packet sent by the receiving end; increase the capacity of the token bucket in the buffer according to the credit packet; when If the size of the data packet to be sent is smaller than the capacity of the token bucket, the data is sent.

Further, after receiving the credit packets sent by the receiving end, according to the proportion of the credit packets with ECN marks, the data packets to be sent are marked with ECN in equal proportion.

Further, after the credit packet arrives, the capacity of the token bucket is increased by 1538B.

Preferably, upon receiving the one credit packet, at least one data packet is sent.

Secondly, the present invention also provides a network congestion control method, which is applied to the receiving end, specifically: receiving the credit request sent by the sending end; sending a credit packet to the sending end; receiving the data flow sent by the sending end.

Further, before sending the credit packet to the sender, the ECT field of the credit packet is set to 1.

Preferably, the sending rate of credit packets is controlled according to the proportion of data packets with ECN marks in the data packets sent from the sending end.

Finally, the present invention provides a network congestion control method applied to switches, which specifically includes: selectively discarding credit packets according to the link capacity limit; performing ECN marking on undiscarded credit packets according to the ratio of the discarded credit packets .

Further, set a data queue and a credit queue at each port of the switch; set the ECN mark threshold K as the tail of the queue; calculate the packet loss rate loss_rate according to the discarded number of the credit packets; The proportion of the package, the ECN marking of the credit package that has not been discarded is specifically: the probability of performing ECN marking on the credit package in the credit queue

P _ECN = σ/loss_rate

Among them, σ is the control factor.

Preferably, the control factor σ has a value of 1.

Compared with the prior art, the technical solution provided by the invention has the following advantages:

1. Credit data packets schedule 1538 tokens instead of one data packet, because one token allows 1 byte transmission, which can avoid credit waste when scheduling data packets.

2. The present invention develops an ECN-based credit rate control mechanism, which can effectively calculate the credit loss rate compared with ExpressPass' credit-based feedback control.

Description of drawings

Fig. 1 is the method schematic diagram of the first embodiment of the present invention;

Fig. 2 is the method schematic diagram of the second embodiment of the present invention;

Fig. 3 is a schematic diagram of the method according to the third embodiment of the present invention.

Detailed ways

The characteristics and exemplary embodiments of various aspects of this specification will be described in detail below. In order to make the purpose, technical solutions and advantages of this specification clearer, this specification will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

First embodiment:

In conjunction with Fig. 1, it is a network congestion control method provided by the first embodiment of the present invention, which is applied to the data sending end, and the method specifically includes:

S11. Setting the token bucket in the buffer before sending the stream data;

Wherein, a token bucket is set in the buffer of the sending end to receive the credit packet sent by the receiving end.

S12. Send a credit request to the receiving end;

When the sending end is ready to send data, it sends a credit request to the receiving end; for ExpressPass, the credit limit is limited to 5% of the link capacity, and the remaining 95% is used to transmit data packets. This mechanism can effectively fully utilize the bandwidth.

S13. Receive the credit packet sent by the receiving end;

After receiving the credit packets sent by the receiving end, according to the proportion of the credit packets with ECN marks in the credit packets, the data packets to be sent are marked with ECN in equal proportion. At the sending end, the data packets corresponding to the ECN-tagged credit packets are also ECN-tagged. In one RTT period, the proportion of received data packets with ECN mark is equal to the average loss rate of credit packets in this RTT.

S14. Increase the capacity of the token bucket in the buffer according to the credit packet;

When the credit packet arrives, increase the capacity of the token bucket in the buffer. ExpressPass uses a 5% rate limit on credits based on the minimum and maximum Ethernet frame size, based on the size ratio of credit packets (84 bytes) and data packets (1538 bytes), respectively.

In this embodiment, a credit packet schedules 1538B of data, not a data packet. A token bucket is set in the send buffer. Once the credit packet arrives, the capacity of the token bucket will increase by 1538B. As long as the size of the data packet to be sent is not greater than the capacity of the token bucket, the data packet can be sent.

S15. When the size of the data packet to be sent is smaller than the capacity of the token bucket, send the data.

The goal of this embodiment is to reduce the delay of data packets with a size smaller than 1MTU (almost all of which are tail packets of short flows and long flows). The main change is in the data sender. First, send as many short streams as possible without the total size exceeding 1538B. Once the cumulative size of these packets exceeds 1MTU, they are sent together. For long tail packets, if the remaining packet size is less than 2MTU but greater than 1MTU, the last two tail packets will be sent out together. However, this poses the challenge that short streams are not necessarily clustered together in the data buffer. Whether it is short flow or long flow, it is sent according to the probability. Long traffic may delay short traffic. Otherwise, in order to send the shortest possible stream, packets need to be rearranged to be transmitted in this way, eg "shortest stream first". This complicates the scheduling process. Therefore, we use tokens instead of credit packets to dispatch data. The credit packet is no longer the number of scheduling data packets, but the data with a size of 1538B. For streams to be sent, the sender sets the token bucket on the egress. The capacity of the token bucket is related to the number of credit packets received. Its corresponding relationship is that the capacity of 1538B corresponds to a credit package. Every time a credit data packet is received, the capacity of the token bucket will increase by 1538B. For data packets exceeding one MTU size, the data packets are still sent according to the size of the largest Ethernet frame (that is, 1538B). A packet that is not the full MTU size can be sent as long as the token bucket has sufficient capacity.

Second embodiment:

In conjunction with FIG. 2, it is a network congestion control method provided by the second embodiment of the present invention, which is applied to the data receiving end. The method specifically includes:

S21. Receive the credit request sent by the sending end;

Once the credit request signal arrives, the data receiver starts delivering credit packets to the data sender.

S22. Send a credit packet to the sender;

Before sending the credit packet to the sender, the ECT field of the credit packet is set to 1.

For credit packets, ExpressPass uses a minimum size 84B Ethernet frame. And each credit packet can schedule the data sender to send the maximum size Ethernet frame (eg 1538B). Therefore, the credit is limited to 5% of the link capacity, while the remaining 95% is used to transmit packets. The receiver also marks the credit packets so that the switch can classify the credit packets and apply rate limits to individual queues.

S23. Receive the data stream sent by the sending end.

According to the proportion of data packets with ECN mark in the data packets sent from the sending end, the sending rate of credit packets is controlled. In each RTT, the receiver adjusts the sending rate of credit data packets according to the credit feedback control.

Third embodiment:

In conjunction with Fig. 3, it is a network congestion control method provided by the third embodiment of the present invention, which is applied to the switch side, and the method specifically includes:

S31. Selectively discard credit packets according to link capacity limitation;

S32. Perform ECN marking on credit packets that have not been discarded according to the discarded credit packet ratio.

Data queues and credit queues are set at each port of the switch; the ECN mark threshold K is set to the tail of the queue; the discarded number calculates the packet loss rate loss_rate according to the number of the credit packets discarded; the discarded credit packets according to the Proportion, performing ECN marking on credit packets that have not been discarded is specifically: in the credit queue, the probability of performing ECN marking on the credit packet

P _ECN = σ · loss_rate

Among them, σ is the control factor. Preferably, the control factor σ has a value of 1.

To balance the three performances of fast convergence, utilization, and fairness, ExpressPass introduces a cyclic credit feedback that dynamically adjusts the sending rate of credit packets. This enables high utilization, fairness and fast convergence. ExpressPass uses a credit packet loss rate as a congestion metric. In this measure, each credit packet carries a sequence number, which is carried and returned by the data packet. The absence of a sequence number indicates that the credit packet has been dropped in the network. When congestion is detected, ExpressPass reduces the rate at which credits are sent to the rate at which credits passed through the bottleneck during the previous RTT.

In the present invention, the credit data packet no longer schedules one data packet, but one or several data packets. Therefore, the sequence numbers of the credit packets cannot be in one-to-one correspondence with the sequence numbers of the data packets. Therefore, the sequence number returned by the data packet is greater than the sequence number of the credit packet. If ExpressPass's method of calculating credit loss rate is used, there will be two counting methods for data packets and credit packets. This complicates the feedback process. Therefore, the ECN marking mechanism is adopted.

The data center switch used in the present invention can well support the ECN marking function. The ECN mechanism uses the ECT field in the IPv4 or IPv6 header to encode whether a packet supports ECN. The package will be marked as ECT(1) if the sender is ECN capable, otherwise it will be marked as ECT(0). Set a single marking threshold K in each exchange queue. When an ECN-capable packet arrives, it will be marked as ECN if the queue length is greater than K.

Therefore, we can use the ECN mark as a parameter to pass the congestion information of the credit packet. On the receiver side, we set the ECT field of all credit packets to ECT(1). In each RTT, the receiver adjusts the rate of credit packets according to the credit feedback control.

Each port of the switch uses a data queue and a credit queue, and they are physically isolated. At the switch, we set the ECN marking threshold K to the tail of the queue. Every time the switch receives ten credit packets, it calculates the packet loss rate loss_rate based on the number of discarded ten credit packets. The probability that the switch marks the last credit packet in the credit queue as ECN is P _ECN =σ·loss_rate. σ is a control factor with a default value of 1. On the sending side, the data packets corresponding to the ECN-tagged credit packets are also ECN-tagged. In one RTT period, the proportion of received data packets with ECN mark is equal to the average loss rate of credit packets in this RTT.

The above descriptions are only examples of the present application, and are not intended to limit the present application. For those skilled in the art, various modifications and changes may occur in this application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims (1)

1. A network congestion control method is applied to a switch, and is characterized in that:

selectively discarding credit packets according to link capacity constraints;

according to the proportion of the discarded credit packets, ECN marking is carried out on the credit packets which are not discarded;

wherein, according to the proportion of discarding the credit packets, ECN marking the credit packets which are not discarded comprises:

setting a data queue and a credit queue at each port of the switch;

setting an ECN marking threshold K as the tail of the queue;

calculating a packet loss rate (loss _ rate) according to the number of discarded credit packets;

the ECN marking of the credit packets that are not discarded according to the ratio of the discarded credit packets specifically comprises: in the credit queue, the probability of ECN marking of the credit packet PECN = sigma/loss _ rate

Wherein, sigma is a control factor; the control factor sigma value is 1.

Images